MKFGO: integrating multi-source knowledge fusion with pretrained language model for high-accuracy protein function...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate protein function identification is crucial for understanding biological mechanisms and advancing drug discovery.
• Experimental methods for protein function determination are often laborious and time-consuming.
• Computational approaches are needed to accelerate protein function prediction.

Purpose of the Study:

• To develop a novel deep-learning method for accurate Gene Ontology (GO) attribute prediction.
• To integrate multi-source biological data for enhanced protein function inference.
• To improve upon existing state-of-the-art protein function prediction methods.

Main Methods:

• Proposed Multi-source Knowledge Fusion for Gene Ontology prediction (MKFGO), a composite deep-learning model.
• Integrated five complementary pipelines utilizing multi-source biological data.
• Developed two core deep-learning components: handcrafted feature representation-based GO prediction (HFRGO) and protein large language model (PLM)-based GO prediction (PLMGO).
• Employed LSTM-attention networks with triplet loss and PLMs for feature extraction and knowledge fusion.

Main Results:

• MKFGO demonstrated superior performance compared to 12 state-of-the-art methods on 1522 nonredundant proteins.
• The combination of HFRGO and PLMGO significantly contributed to MKFGO's accuracy.
• Complementary insights from protein-protein interaction, GO term probability, and gene sequences further enhanced prediction.

Conclusions:

• MKFGO offers a powerful and accurate approach for predicting protein functions using deep learning.
• The method's strength lies in its multi-source data integration and decision-level knowledge fusion.
• The developed model and source code are publicly available for research use.